Hyperspectral Deep Learning Mineral Type Identification Based on Samples Generated from Ground Object Spectral Library

Abstract

Despite the success of deep learning-based methods for hyperspectral mineral classification, it depends on large scale labelled data and for new data outside training set, it will result in misjudgment. And there are some problems, such as over fitting of local features and low generalization ability. To address this problem, we propose a method of build sample database based on the United States Geological Survey (USGS) spectral library. On this basis, mineral type recognition of hyperspectral data is carried out in the deep learning method. The mineral mapping experiments selected the AVIRIS hyperspectral data of the Cuprite mining area in Nevada. The overall accuracy reached 76.41 %. This method improves the generalization ability when the method is used in data outside training set and breaks through the dependency on labelled data. The efficiency and accuracy of mineral type identification are improved while maintaining high accuracy.